摘要： 经典竞争聚集(CA)算法在聚类时对于样本中的少量已知信息没有加以利用,但这些信息往往需要应用到整个聚类过程中。此外,在相似度度量函数的选择上CA 算法使用常见的欧氏距离,然而欧氏距离仅适用于团状数据,制约了算法的应用范围。针对上述问题,通过引入具备半监督学习能力的半监督项对隶属度矩阵进行增强,利用聚类中心和中心邻近的点组成空间,把样本点与该空间的距离替代欧氏距离作为新的相似度度量标准,并给
出判断聚类中心能否合并的阈值参数,最终得到半监督空间化CA 算法。通过在人造图像和真实图像上的分割结果表明,该算法能够更准确地获取聚类类别数以及更好的聚类效果。
关键词:
竞争聚集算法,
相似度度量函数,
欧氏距离,
半监督,
空间距离,
阈值参数
Abstract: Classic Competitive Agglomeration(CA) algorithm fails to take into account of the information about a few
samples which are known and important during the process of cluster. Moreover,competitive agglomeration chooses
Euclidean distance as a similarity metric function. But,the distance is more applicable when the distribution of the data points is spherical. This restricts the scope of its application. In order to solve these problems,the semi-supervised entry with the ability to learn is introduced to enhance membership matrix. And a distance from a sample to the spaces is used to instead of Euclidean distance,that each of them is composed of one cluster center and its nearest points. A threshold parameter about similarity of cluster centers is introduced to algorithm as the judgment for merging. The semi-supervised spatial distance competitive agglomeration is proposed. Two sets of experiments using artificial image and real images are operated,and the results show that the proposed algorithm has greater ability to get right cluster number,and gets better
clustering results.
Key words:
Competitive Agglomeration ( CA ) algorithm,
similarity metric function,
Euclidean distance,
semisupervised,
spatial distance,
threshold parameter
中图分类号: